Fatykhich commited on
Commit
16e685f
·
verified ·
1 Parent(s): f9e0471

Add MuViSeg trained heads (ACCV 2026)

Browse files
README.md ADDED
@@ -0,0 +1,101 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: mit
3
+ library_name: pytorch
4
+ pipeline_tag: image-feature-extraction
5
+ tags:
6
+ - segment-matching
7
+ - multi-view
8
+ - correspondence
9
+ - 3d-vision
10
+ - mast3r
11
+ - vggt
12
+ ---
13
+
14
+ # MuViSeg checkpoints
15
+
16
+ Trained heads for **MuViSeg: Multi-View Segment Correspondences from Dense
17
+ Geometry Priors** (ACCV 2026).
18
+
19
+ - Code: https://github.com/muviseg/muviseg-codeRelease
20
+ - Project page: https://muviseg.github.io
21
+ - Paper: https://arxiv.org/abs/2607.17938
22
+
23
+ Given two or more images of the same scene plus class-agnostic segmentation masks
24
+ for each, MuViSeg predicts which segments correspond to the same physical object.
25
+ A frozen 3D foundation-model backbone (MASt3R or VGGT) supplies dense descriptors,
26
+ masked average pooling gives one descriptor per segment, and a lightweight
27
+ trainable head (LightGlue-style attention with a DoubleSoftmax matcher) predicts
28
+ the assignment, with a dustbin for unmatched segments.
29
+
30
+ ## These are head-only checkpoints
31
+
32
+ The frozen backbone is stripped and reloaded separately at inference time, so each
33
+ file is a few MB rather than a few GB. Each holds the trainable head plus `epoch`,
34
+ `global_step`, `best_val_ma` and the validation `metrics`. **They cannot resume
35
+ training** — that needs the optimizer state of a full checkpoint.
36
+
37
+ The backbones are not hosted here; they belong to their upstream authors:
38
+ VGGT-1B from [facebook/VGGT-1B](https://huggingface.co/facebook/VGGT-1B), and
39
+ MASt3R from the [MASt3R release](https://github.com/naver/mast3r#checkpoints).
40
+
41
+ | file | model | trainable params | size |
42
+ |---|---|---|---|
43
+ | `segmast3r_lg_v2/best.pth` | SegMASt3R + LG v2 | ~0.8 M | 3.3 MB |
44
+ | `segvggt_dpt/v3-001/best.pth` | SegVGGT-DPT, pairwise | ~10.7 M | 50 MB |
45
+ | `segvggt_dpt/joint-001/best.pth` | SegVGGT-DPT, joint attention over N frames | ~11.0 M | 49 MB |
46
+ | `segvggt/best.pth` | SegVGGT single-layer (ablation) | ~0.8 M | 5.5 MB |
47
+ | `segvggt/step_0140000.pth` | SegVGGT single-layer, step 140k | ~0.8 M | 5.5 MB |
48
+
49
+ The single-layer row reported in the paper uses `segvggt/best.pth` on Replica and
50
+ `segvggt/step_0140000.pth` on Virtual KITTI 2, which is why both are published.
51
+
52
+ ## Results
53
+
54
+ Overall AUPRC / R@1 / R@5, on the frozen benchmark pair lists shipped with the code.
55
+
56
+ | model | Replica | Virtual KITTI 2 |
57
+ |---|---|---|
58
+ | SegMASt3R + LG v2 | **84.5** / 77.5 / 93.6 | **81.2** / 72.7 / 92.8 |
59
+ | SegVGGT-DPT (pairwise) | 79.7 / 74.1 / 90.1 | 51.9 / 42.9 / 67.6 |
60
+ | SegVGGT-DPT Joint (N=4) | 81.6 / 76.5 / 90.6 | 52.8 / 43.2 / 66.8 |
61
+ | SegVGGT single-layer | 79.2 / 73.3 / 89.6 | 51.5 / 42.1 / 66.3 |
62
+
63
+ MASt3R + LG v2 dominates at small viewpoint change; joint multi-frame attention
64
+ narrows the gap at large rotations. The code repository's `docs/reproduction.md`
65
+ states which rows are exactly reproducible and which are not — read it before
66
+ comparing against the paper.
67
+
68
+ ## Usage
69
+
70
+ ```bash
71
+ git clone https://github.com/muviseg/muviseg-codeRelease.git
72
+ cd muviseg-codeRelease
73
+ uv sync
74
+ bash setup_third_party.sh
75
+ uv run python scripts/download_checkpoints.py --all --verify
76
+ ```
77
+
78
+ That places every file where the shipped configs expect it. To fetch one directly:
79
+
80
+ ```python
81
+ from huggingface_hub import hf_hub_download
82
+
83
+ path = hf_hub_download("muviseg/muviseg", "segvggt_dpt/v3-001/best.pth")
84
+ ```
85
+
86
+ ## License
87
+
88
+ MIT, as the code. Third-party components (MASt3R, VGGT, SegMASt3R, LightGlue,
89
+ RoMa, SAM 2, FastSAM) remain under their own licenses.
90
+
91
+ ## Citation
92
+
93
+ ```bibtex
94
+ @inproceedings{muviseg2026,
95
+ title = {MuViSeg: Multi-View Segment Correspondences from Dense Geometry Priors},
96
+ author = {Fatykhoph, Denis and Akhtyamov, Timur and Pakulev, Konstantin
97
+ and Devchich, German and Ferrer, Gonzalo},
98
+ booktitle = {Proceedings of the Asian Conference on Computer Vision (ACCV)},
99
+ year = {2026}
100
+ }
101
+ ```
segmast3r_lg_v2/best.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:93e76745d1de5f39fc699ccae8fe8717e5f1e2bcef450a8e33aca3e6cf8d8117
3
+ size 3287523
segvggt/best.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:2aa73caad3a4939a57ee74e458d3aef29e41760c99ba9d70a78c0a1654d26143
3
+ size 5479203
segvggt/step_0140000.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d39d81153d601417fc67703c57b71171f15b5351a99f73f72ff5fbe2d38e11cb
3
+ size 5483387
segvggt_dpt/joint-001/best.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ef93791cc73edd3965e26f30952917aa1099638ee129c6d45a4aa0bc72c583b7
3
+ size 48775717
segvggt_dpt/v3-001/best.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4ad62ab59c847bce37769f646fa0ccb3117bdaed83e914f44d7bcc3529ff1e18
3
+ size 49547769